[Core] Optimize SPMD architecture with delta + serialization optimization (#7109)

tests/distributed/test_chunked_prefill_distributed.py

@@ -6,6 +6,8 @@ pytest test_chunked_prefill_distributed.py
 ```
 """
 
+import os
+
 import pytest
 
 from vllm.utils import cuda_device_count_stateless
@@ -30,6 +32,11 @@ def test_models(
     model: str,
     distributed_executor_backend: str,
 ) -> None:
+    if model == "meta-llama/Llama-2-7b-hf" and distributed_executor_backend == "ray":  # noqa
+        assert distributed_executor_backend == "ray"
+        # test ray adag
+        os.environ['VLLM_USE_RAY_SPMD_WORKER'] = "1"
+        os.environ['VLLM_USE_RAY_COMPILED_DAG'] = "1"
 
     dtype = "half"
     max_tokens = 5